Action Recognition of Excavator by Deep Learning Using Simulated Training Data

Abstract

Measuring and recording an action performed by construction machinery is a very effective task for improving productivity of construction sites. However, measuring and recording the action by construction machinery is time consuming and expensive because the construction site managers have to observe and record data manually. Therefore, it is important to automatically recognize the action of construction machinery. Action recognition is achieved with high performance on humans using deep learning techniques but those approaches require large amounts of training data. There is no data set for action recognition of construction machinery. Therefore, the proposed method uses training data generated from a simulator. In this study, action recognition is performed for excavator that is most commonly used at construction sites. Experiments were conducted with a remote control excavator in laboratory conditions.